The observer in a complex system: why continuous ingestion follows

The observer in a complex system: why continuous ingestion follows

Start with a plain fact about measurement. To know the temperature of a room, you put a thermometer in it. The thermometer exchanges heat with the room. A large glass thermometer in a small volume of water will cool the water perceptibly while it reads it. The reading is real, but it is a reading of a system that now includes the thermometer. There is no version of this measurement taken from outside, because there is no outside: the instrument is a component, coupled to the state it reports.

Scale that up and the fact does not go away; it just gets easier to ignore. A weather bureau publishes a forecast, and airlines rebook flights, which changes fuel loads, which changes the very air traffic the next forecast will describe. A bank supervisor requests a liquidity return, and banks start managing their balance sheets to the number the return asks for, rather than to whatever the number was originally meant to capture. The observation was supposed to be a window. It turns out to be a valve.

Systems theory has a name for this: the observer is not outside the system observed. It is a component whose state is coupled to the system's state, drawing information out through some channel and, by the act of drawing, changing what happens next. Stop observing and the coupling breaks. The model that remains describes a system the observer is no longer part of — which is a polite way of saying it describes a system that no longer quite exists.

Where the idea comes from

Norbert Wiener's cybernetics, formulated in 1948, put feedback loops at the centre of control: a system regulates itself by sensing its own output and correcting toward a target. That already implies an observer inside the loop, but Wiener's observer was mostly a thermostat — simple, mechanical, easy to bracket off as "just the sensor." Heinz von Foerster, working at the Biological Computer Laboratory through the 1960s and 1970s, pressed the question Wiener had left comfortable: where does the observer of the whole cybernetic system stand? If the observer must itself be modelled to close the loop, it cannot stand outside the model. Von Foerster called this second-order cybernetics — the cybernetics of observing systems, rather than of observed ones. Humberto Maturana and Francisco Varela extended the idea into autopoiesis, treating living systems as defined by the very processes that observe and maintain them. Ross Ashby's law of requisite variety gave the whole programme a quantitative edge: a regulator can only control a system whose variety it can match, which means the observer's capacity, not just its presence, is part of the system's dynamics. The problem being solved throughout was self-reference — how a model of a system can include the thing doing the modelling without collapsing into contradiction.

The Hawthorne studies at Western Electric, 1924 to 1932, are the case every methods course eventually reaches for, and reasonably so. Productivity rose under nearly every lighting condition tested, including a reversion to the original levels. The details have been re-litigated and much of the folklore about the effect is overstated. What survives the re-litigation is the point that mattered to systems theorists: measuring workers is an intervention on workers. The observation is not a neutral tap on a pre-existing signal. It is a new input to the system, arriving through a channel — attention, in that case — that the system responds to.

The turn

None of this, so far, has anything to do with machine learning. It is a claim about instruments, regulators, and observed populations — thermometers, supervisors, factory floors. But the claim generalises cleanly to any system that takes in information about a world and acts, or is later used to act, on the basis of what it took in. That generalisation is where the lineage from Large Language Model to Large World Model to Large Universe Model sits, and it is worth being precise about why.

A Large Language Model observes once. A corpus is scraped, filtered, frozen at some cutoff; the world then continues without it. Whatever coupling existed between the model's training data and the world at collection time starts decaying the moment collection stops, and nothing in the model's outputs decays alongside it — a query in year three gets an answer calibrated, in tone and confidence, as though it were still year zero. In von Foerster's terms, this is an observer that took one reading and then physically left the room. The reading is preserved. The coupling is gone.

A Large World Model rejoins the loop, but only for a while. Sensors run, a scene is perceived, the feedback loop between action and perception closes for the length of an episode — a robot manipulating objects on a table, a vehicle navigating a junction. Real coupling, genuinely bounded. When the episode ends, the observer steps back out, exactly as it did after the corpus scrape, just on a shorter and more forgiving cycle.

A Large Universe Model is the position defined by refusing that step-out. Streams stay open; beliefs are held with provenance — where a claim came from, when, at what confidence — and are revised as new observation contradicts old. This is not simply a bigger pipe. It is a semantic claim. A frozen belief does not mean what it once meant, because the thing it referred to has moved on; only continuous coupling keeps a stored claim and the state of the world pointing at each other. Once observation is understood, correctly, as a standing relation rather than a one-off act, the intake axis acquires a floor and a ceiling by logical necessity rather than by engineering ambition. The floor is a single historical sample. The ceiling is unbroken coupling, everywhere the system can reach, with revision built in. There is no relation between an observer and evidence that sits further inside the loop than being always inside it.

The objections that actually bite

The first objection is that perpetual observation is not free. Every sensor draws power, every channel carries load, and — Hawthorne again — every act of watching perturbs the thing watched. An observer that never stops may be an observer that never stops interfering, and selective, episodic observation might be the better systems design, not merely the cheaper one. This is correct as stated, and it narrows the claim usefully: the terminal position does not require saturation. It requires that the decision about what to observe stay live, rather than being fixed once, at scrape time, by whoever built the corpus. A system sampling one stream at a hundredth of a hertz and another at a thousand hertz is still inside the loop; a frozen billion-token corpus is not, no matter its size. Interference is a budget to manage from inside the loop. It is not a reason to leave it.

The second is more serious, and deserves to be granted in full. Continuous observation does not guarantee continuous understanding. A system can stream data forever while holding the same stale categories, thresholds, and causal assumptions it started with, applied to a world that has since restructured underneath them. Drift in concepts is not repaired by drift-free data. This is the honest limit of the whole argument: continuous intake is necessary for revising an ontology, not sufficient for it. But note what a frozen system cannot even do — it cannot notice its own categories have failed, because it has no post-cutoff surprise to notice with. Provenance-tagged, revisable belief is the mechanism that turns anomaly into category change; continuous intake is what makes that anomaly arrive at all. The further work, of actually revising the categories, happens after the intake axis has already reached its terminus.

The third objection tries to find a fourth rung: counterfactual intake, where a system intervenes to generate observations that would not otherwise exist, or multi-observer intake, where several observers negotiate a shared frame. Both matter enormously. But intervention is a policy for creating new streams, which are then either observed continuously or not — it sits inside the terminal position rather than beyond it. Multi-observer negotiation is a question of how many loops there are and how they reconcile, which is a scaling and trust problem, not a new relation between an observer and its evidence.

If nothing is outside the system, then nothing can be verified from outside it either — so every model is equally compromised, and the only honest response is to watch everything, all the time.

This is the misreading to disown, explicitly. Observer coupling is a measurable, usually bounded, often small effect — the thermometer cools the water a little, not infinitely — not a licence for relativism about which models are better. And the claim was never "watch everything." It was that the choice of what to watch must stay open rather than being frozen once, at the moment a corpus is built. Total surveillance is a different proposal, and a worse one, with its own costs that have nothing to do with this argument.

What this establishes, stated no larger than it should be: intake sits on an axis with a real floor and a real ceiling, because observation is a position within a system rather than a vantage outside it, and there is no position more inside than permanent membership in every reachable loop. What it does not establish: that continuous intake yields correct belief, that interference costs are always worth paying, or that representational understanding follows automatically once the streams stay open. Those are separate arguments, resting on this one, not settled by it.